Intelligence Got Cheaper, Money Got Dearer: Three AI Labs Cut Frontier Prices in 48 Hours, Japan's 10-Year Yield Hit a 30-Year High — and How to Re-Run Your AI Business Case for Japan

The first cut came from xAI. On Monday, September 21, it shipped Grok 4.7 at $2 per million input tokens and $6 per million output tokens — about 80% below the $10/$50 price that had marked the top of the market. On Tuesday, September 22, Anthropic released Claude Opus 5.5 at $4 and $20, down from $5 and $25 for Opus 5. Cache reads, the price of re-using a long prompt or document the model has already seen, fell 60%, from $0.50 to $0.20 per million tokens. Anthropic says Opus 5.5 runs more than 30% faster and matches its larger Fable 5.1 model on most tasks, and it is available through AWS, Google Cloud and Microsoft Azure.

About ninety minutes later, OpenAI answered with two models. GPT-6 Sol costs $2 per million input tokens and $10 per million output, and GPT-6 Luna costs $0.10 and $0.50. OpenAI says both are 50% cheaper than the GPT-5.6 series they replace, both read up to 1.05 million tokens at once, and repeated input served from cache is discounted by 90%. Luna costs one hundredth of OpenAI's top model, GPT-6 Astra, at $10 and $50. In practical terms, a routine task such as sorting incoming emails or tagging product listings now costs a fraction of a cent per item.

The timing matters. Opus 5.5 is Anthropic's first model since its CEO, Dario Amodei, published “We Must Pace the Frontier” on September 12, the essay we covered last week. Some readers took that essay as a signal that AI progress would stall. This week shows what the labs actually agreed to: more caution on the most dangerous capabilities in training, not fewer releases and not higher prices. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. For a company buying AI, the frontier did not slow down. It became cheaper to use.

Two days after the price cuts, the bond market moved the other way. On Thursday, September 24, the yield on Japan's 10-year government bond rose 8 basis points to 3.06%, the highest since August 1996. The 30-year yield reached 4.12%. In the United States, the 10-year Treasury yield passed 5.1%, its highest since 2007. Brent crude traded above $105 a barrel. Japan's 10-year yield has more than tripled in two years and has roughly doubled since Prime Minister Sanae Takaichi took office last October on a platform of heavy public spending.

This came six days after the Bank of Japan lifted its policy rate to 1.25%, the move we covered last week. The yen did not strengthen on the higher yields. The dollar closed near ¥159 on September 24, about 6% stronger against the yen than a year earlier. For a foreign company, a weak yen makes Japanese salaries, offices and acquisitions cheaper in dollars or euros. It also makes every dollar-priced cloud and AI bill more expensive for a Japanese customer.

The stock market did not panic. The Nikkei 225 rose 495 points to 65,513.99 on the same Thursday, led by AI and semiconductor stocks, and it was one of the few Asian markets to gain that day. That is the pattern to understand: investors still pay for AI growth, but the cost of the money behind every long project has gone up. A Japanese company that approves an AI budget this winter will do it with a 3% government yield on the finance team's screen, a level no one in that team has seen in their working life.

An AI project has three kinds of cost. The first is the model bill: the tokens. The second is the build: connecting the model to the company's systems, cleaning the data, testing the output in Japanese, and training the staff who will use it. The third is time: the months between signing the budget and seeing the savings. This week's price cuts reduce only the first. In most projects we see in Japan, the build is the largest line, and it is mostly people — engineers, project managers and the client's own staff. Those costs are rising, not falling: Japan raised its minimum wage 4.9% this year, and engineering rates follow.

Here is a simple example. A Japanese retailer plans an AI assistant for its customer-service desk. The build costs ¥24 million, the model bill is ¥400,000 a month, and the assistant saves ¥2.4 million a month in staff time. Net saving is ¥2 million a month, so payback takes twelve months. Apply this week's 20% list-price cut to the model bill and the net saving rises to ¥2.08 million — payback drops by about two weeks. Now cut the build to ¥16 million by starting with one product line instead of five. Payback falls to eight months. The build decides the payback, not the tokens.

Higher yields make that difference matter more. When money was almost free, a Japanese company could accept a project that paid back in two or three years. With the 10-year yield at 3.06% and the BOJ at 1.25%, finance teams will ask for payback inside one fiscal year, and ringi approval — the circulated sign-off that most Japanese budgets need — will move faster for small, fast projects than for large platforms. There is one currency detail worth noting: the 20% cut on Opus and the 50% cut on GPT-6 more than cover the roughly 6% the yen lost against the dollar in the past year, so a dollar-priced model bill is now cheaper in yen than it was twelve months ago.

On the same Tuesday as the model launches, Cisco's threat-intelligence unit, Talos, disclosed a piece of Windows malware it calls CLOSEDQUORUM and described it as the first reported implant with fully autonomous AI command and control. Normal malware waits for instructions from a human operator. CLOSEDQUORUM collects details about the infected computer, sends them as prompts to four commercial AI models — DeepSeek, Qwen, Mistral and Google Gemini — and lets them vote. It then carries out whichever action wins, aimed at stealing passwords and cryptocurrency wallets. Talos found no confirmation that it has been used against victims yet, and it has released a toolkit, CAIRN, to help defenders find malware that calls AI services.

The business lesson is not that AI is dangerous in general. It is that the cost of running an attack is falling at the same rate as the cost of running a customer-service assistant. When a model call costs a fraction of a cent, an attacker can afford to let software make every decision, around the clock, with no one at the keyboard. Small and mid-size companies, which in Japan make up the large majority of businesses and often lack a dedicated security team, are the obvious targets.

For a company rolling out AI in Japan, this means two practical changes. First, watch which of your computers talk to AI services. A laptop in the accounts team that suddenly sends prompts to four different model providers is a warning sign that standard antivirus may miss. Second, put AI security in the same budget request as the AI project. Japanese buyers already ask what an agent can do without a human. Next year they will also ask how you would notice if someone else's agent were running inside their network.

First, re-quote every AI proposal you have open in Japan this week. Any quote built on September 21 prices is now too high on the model line, and a Japanese buyer who reads the news will notice before you do. Show the new model cost, but lead with the build and the payback in months, because that is where the finance team will look. If a proposal pays back in more than twelve months, cut its scope until it does not — one department, one product line, one language pair — and put the rest in a second phase.

Second, route work to the right model instead of sending everything to the most expensive one. Bulk, repetitive work — sorting enquiries, tagging products, first-draft translation between Japanese and English — belongs on a low-cost model such as GPT-6 Luna, which costs one twentieth of Sol. Keep Opus 5.5 or Sol for the tasks where a wrong answer costs money: contract review, customer-facing Japanese copy, anything a manager signs. And write price pass-through into every contract. Prices fell 20% to 50% in one day; a three-year fixed-price AI deal signed last month is already overpriced, and the next cut may come with Sonnet 5.5 and Haiku 5.5 in the coming weeks.

Third, move now. Japanese budgets for the fiscal year starting in April are drafted between October and January, and this winter they will be drafted with a 3% government bond yield in view. A small project that pays back in eight months and includes a security line will clear the ringi. A large platform that pays back in thirty months will wait. At Medusa Japan, this is the work we do with foreign companies entering Japan: we size the first AI project so it pays back inside one Japanese fiscal year, choose the model mix, and write the proposal in the form a Japanese approval process expects.